AI Copywriting: When It Works and When It Does Not

The short answer

AI copywriting works best when the thinking is already done and the job is expression: turning a clear brief into drafts, variants, shorter versions or channel-specific formats. It works worst when the thinking is the job: deciding what to say, finding the angle nobody else has, telling a story only one person lived, or making a claim where one wrong word creates legal risk.

Most disappointment with AI copy comes from giving it the second kind of job and judging it on the first kind of skill. The draft is fluent, so it looks finished; the problem is that it says nothing, or says something untrue.

This article is about when to use AI at all. For which structure to use once you do, see AI copywriting frameworks.

Where it works, where it needs help, where it does not

Copywriting jobs by AI suitability

Job	Suitability	Why
Variants of approved copy (headlines, subject lines, ad assets)	Works well	The message is decided; AI supplies range quickly, and testing picks winners
Compressing or expanding approved copy for another channel	Works well	Meaning is fixed by the source; the task is fitting length and format
Responsive search ad assets	Works well with rules	Needs many standalone headlines and descriptions within strict character limits, which suits generation plus human selection
First drafts from a rich brief	Works well	When the brief carries the angle, proof and customer language, the draft is mostly assembly
Product descriptions from real specs	Works with checks	Efficient at volume, but every factual detail must match the spec
Website positioning and homepage headline	Needs heavy human input	The value is the decision about what you stand for; AI can offer options, not choose
Thought leadership in a named person's voice	Needs heavy human input	Readers are buying the person's judgement; AI can shape it, not supply it
Humour and brand wit	Rarely works	Models default to safe, familiar jokes; distinctive humour depends on taste and timing
Personal stories and testimonials	Should not be generated	Inventing experiences or customer words is fabrication, and fake reviews are unlawful in the UK
Regulated claims (financial, health, legal)	Human-led only	Precision and evidence requirements are high; AI may help check, not originate
Crisis or apology communications	Human-led only	Tone errors are costly and the facts change hour to hour

The four-question test

For any copy job, answer these four questions. Score yes = 2, partly = 1, no = 0.

1. • Is the input rich? Is there a brief with the angle, proof, audience and customer language already decided?
2. • Are the claims checkable? Could a person verify every factual statement against a source in under a minute?
3. • Is volume or variation useful? Would more options, versions or formats genuinely help (testing, channels, languages)?
4. • Is the cost of an error low? If a wrong or off-tone line went out, would the damage be small and fixable?

7 or 8: let AI draft, with a normal review. 4 to 6: AI can assist (options, structure, editing), but a person writes the key lines and checks everything. 0 to 3: a person writes it; AI may help check spelling, length or consistency.

The test in practice

Illustrative examples, not client work.

Twenty Google Ads headline variants for an existing, approved offer. Input rich (yes, 2), claims checkable (yes, 2), variation useful (yes, 2), low error cost (partly, 1: ads cost money, but can be paused). Score 7: AI drafts; a person selects and checks character limits. Google's responsive search ads use up to 15 headlines of up to 30 characters each, so generation plus selection is a natural fit.

A new homepage headline after a repositioning. Input rich (partly, 1: the positioning is still being argued), claims checkable (yes, 2), variation useful (partly, 1), low error cost (no, 0: it is the first thing every visitor reads). Score 4: AI can generate a long list of options to react to, but the choice and final wording are human.

A service page for a regulated financial product. Input rich (yes, 2), claims checkable (partly, 1: the rules around wording are strict), variation useful (no, 0), low error cost (no, 0). Score 3: a person writes it, with the compliance owner.

The two failures to watch for

Invented specifics. Asked to make copy "more specific", a model will add numbers, names and quotations that look real. In UK advertising, the ASA expects marketers to hold evidence for objective claims before publishing, so an invented "saves 12 hours a week" is not just embarrassing; it is a compliance problem. Supply the specifics yourself or remove them.

Leaked context. Pasting customer lists, contracts or personal data into a general AI tool to "personalise" copy is processing personal data. The ICO's position is that there is no AI exemption from data protection law. Use anonymised or public information unless your tool and contract are approved for it.

Where the human effort goes when AI does the drafting

The effort moves, it does not disappear. It moves to the brief (deciding what to say), to selection (choosing among options) and to verification (checking every claim). Expect to spend more time on those three than before, and much less on first drafts. Skip them and you get more copy that performs no better.

FAQ

Is AI copy worse than human copy?

Not inherently. For variants and channel adaptations from a strong brief, it is often as good and far faster. It is worse where the value lies in a decision, a lived experience or a distinctive taste that the model does not have.

Can AI write customer testimonials if we base them on real feedback?

Do not generate testimonials. Quote real customers with their permission. A generated version presented as a customer's words is a fabricated review, and fake reviews are banned in the UK under consumer protection law.

Should AI copy be A/B tested differently?

No, but it makes testing cheaper to feed. Generate several variants, choose the most distinct ones for the test, and keep the winner as an exemplar for the next round of generation.

How do I stop AI copy from exaggerating?

Give it the proof it may use and instruct it to make no claim beyond that proof. Then check every number and superlative against your evidence before publishing.